*m r DAVID W. TAYLOR NAVAL SHIP RESEARCH AND DEVELOPMENT CENTER Bethesda, Md. 20084 CONTROL THEORY WITH APPLICATIONS TO NAVAL HYDRODYNAMICS THE FIRST DAVID W. TAYLOR LECTURES APRIL 1972 by Professor Dr. Reinier Timman Technische Hogeschool Delft, Netherlands Notes by Thomas J. Langan APPROVED FOR PUBLIC RELEASE: DISTRIBUTION UNLIMITED December 1975 Report 4397 J MAJOR DTNSRDC ORGANIZATIONAL COMPONENTS DTNSRDC COMMANDER 00 TECHNICAL DIRECTOR 01 OFFICER-IN-CHARGE CARDEROCK 05 SYSTEMS DEVELOPMENT DEPARTMENT 11 OFFICER-IN-CHARGE ANNAPOLIS 04 UNCLASSIFIED SECURITY CLASSIFICATION OF THIS PAGE (When Dala Entered) REPORT DOCUMENTATION PAGE READ INSTRUCTIONS BEFORE COMPLETING FORM I. REPORT NUMBER 4397 2. GOVT ACCESSION NO 3. RECIPIENT'S CATALOG NUMBER 4. TITLE (and Subtitle) CONTROL THEORY WITH APPLICATIONS TO NAVAL HYDRO- DYNAMICS: THE FIRST DAVID W. TAYLOR LECTURES — APRIL 1972 5. TYPE OF REPORT & PERIOD COVERED 6. PERFORMING ORG. REPORT NUMBER 7. AUTHORfs) 8. CONTRACT OR GRANT NUMBERfs; Reinier Timman (Technische Hogeschool, Delft, Netherlands) 9. PERFORMING ORGANIZATION NAME AND ADDRESS Work Unit 4-1500-001-43/1509 CONTROLLING OFFICE NAME AND ADDRESS David W. Taylor Naval Ship R&D Center Bethesda, Md. 20084 12. REPORT DATE December 1975 '3. NUMBER OF PAGES 75 14. MONITORING AGENCY NAME a ADDRESSff/ different from Controlling Office) 15. SECURITY CLASS, (of this report) UNCLASSIFIED 16. DISTRIBUTION ST ATEM EN T (of this Report) Approved for Public Release: Distribution Unlimited 17. DISTRIBUTION STATEMENT (of the abstract entered In Block 20, II different from Report) 18. SUPPLEMENTARY NOTES Material contained in this report was presented in four lectures by Dr. Timman. Collation and technical editing was done by Dr. T. J. Langan, DTNSRDC Code 1552. 19. KEY WORDS (Continue on reverse side It necessary and Identify by block number) Modern control theory Stochastic systems Calculus of variations Kalman-Bucy filter solution Method of dynamic programming 20. ABSTRACT (Continue on reverse side If necessary and Identify by block number) The lectures present an introduction to modern control theory. Calculus of variations is used to study the problem of determining the optimal control for a deterministic system without constraints and for one with constraints. The method of dynamic programming is also used to solve the unconstrained control problem. Stochastic systems are introduced, and the Kalman-Bucy filter is derived. DD 1 JAN 73 1473 EDITION OF 1 NOV 65 IS OBSOLETE S/N 0102-014-6601 UNCLASSIFIED SECURITY CLASSIFICATION OF THIS PAGE (IVhen Data Entered) UNCLASSIFIED -H.IJWITY CLASSIFICATION OF THIS FAGEf UNCLASSIFIED SECURITY CLASSIFICATION OF THIS P AGE(When Dele Entered) TABLE OF CONTENTS Page ABSTRACT 1 INTRODUCTION 1 THE OPTIMAL CONTROL PROBLEM 6 RELATION TO DYNAMIC PROGRAMMING 23 CONSTRAINTS ON THE CONTROL AND STATE VARIABLES 26 STOCHASTIC SYSTEMS 40 THE KALMAN-BUCY FILTER 53 LIST OF FIGURES 1 — Geometry of the Proof 17 2 — Constant Cost Fronts 23 3 — Constrained Variables 28 4 — Optimal Trajectory 36 5 — Switching Curve 36 6 — Time Fronts 39 7 — Stochastic Control System 41 8 — Optimal Filter 62 PREFACE The David W. Taylor Lectures were initiated as a living memorial to our founder, in recognition of his many contributions to the science of naval architecture and naval hydromechanics. His systematic investiga- tion of resistance of ship hulls is universally known and used, but of equal importance was his use of hydrodynamic theory to solve practical problems. Many of the experimental techniques which he pioneered are still in use today (for example, the use of a spherical pitot tube for exploring the structure of a wake field). The system of mathematical lines developed by Taylor was used to develop many designs for the Navy long before the computer was invented. And perhaps most important of all, he established a tradition of applied scientific research at the "Model Basin" which has been carefully nurtured through the decades, and which we treasure and protect today. These lectures were conceived to support and strengthen this tradition. We will invite eminent scientists in fields closely related to the Center's work to spend a few weeks with us, to consult with and advise our working staff, and to give lectures on subjects of current interest. It is most fitting that Professor Reinier Timman, mathematician and philosopher, initiate this series. He has long been a friend and on several occasions has used the Center for a retreat, to his benefit and ours. He has inspired and advised our staff and cooperated in our work. His students at Delft have made leading contributions to the development of modern naval hydrodynamics. Professor Timman' s belief that mathe- matics can contribute powerfully to our technology is much in the David Taylor tradition. We are honored that he agreed to give the first in this David W. Taylor Lecture Series. W. E. CUMMINS FOREWORD It is great honor to me to be invited to give the first in the series of David W. Taylor Lectures. My associations with the Model Basin date from a long time ago, and a visit to the United States is for me not a real visit unless I have the opportunity to taste once more the stimulating atmosphere which not only gives the Model Basin an out standing place in hydrodynamical research but also acts as a breeding ground where nearly all outstanding people in the field passed an essential period in their lives. So I am extremely grateful to have been given the opportunity once more to spend some time at this most interesting place and to participate in its work. I wish to express my gratitude to Justin McCarthy who originated the idea of the lectures and to all other friends who made this period a success. In particular, I am pleased that Dr. Langan, whom I used to know as a promising undergraduate student, did a fine job in editing the lectures. R. TIMMAN vi ABSTRACT The lectures present an introduction to modern control theory. Calculus of variations is used to study the problem of determining the optimal control for a deterministic sys- tem without constraints and for one with constraints. The method of dynamic programming is also used to solve the unconstrained control problem. Stochastic systems are intro- duced, and the Kalman-Bucy filter is derived. INTRODUCTION Optimal control theory is involved with the great human effort to control or influence processes of one type or another. The objectives and criteria for the performance of a physical system may be diffused or defy tractable analysis in many situations, but the basic concepts on which to proceed have been established in control theory. One first considers a system and a process through which the state of the system is changing in time; in other words, some action or motion of the system takes place in time. This behavior of the system is described by a set of time-dependent variables x (t) = (x.. , . . . , x ) which are called the state variables. In addition to the state of the system, one also considers controls by which the process in question can be influenced. These controls are represented by a set of variables u (t) = (u-t (t) , . . . , u (t)) which are called the control variables. 1 m At a certain instant in time, say t„, the state of the system is li known to be x n . If an analysis of the system is to be performed, a sys- "0 tem of equations must be specified which predict the state for t > t. and for a given control function _u. ' These equations are called the dynamic equations for the system; they may take the form of an ordinary differential equation X (t) = f (t, x, u) or a difference equation X - = f (t , X , u ) n+1 n n n They might even take the form of an integro-dif ferential-dif ference equation or a time delay equation, but they cannot take on a form such that, the solution at some time t is dependent on the solution in the future, t > t . The dynamic equations must reflect this principle of nonanticipation. One does not violate this principle by choosing a control in anticipation of the future and thus influencing the future state of the system based on estimated future information; in fact, the choice of such a control is actually based on the history of the state of the system available at the time of the choice. If no further specification of system performance is given, every control function which yielded a physical realizable state of the system for t > t„ would be a solution to the control problem. One can have a meaningful control problem only if there is a desired objective, a goal to be achieved by the process. Moreover, it is not sufficient merely to have a goal; there must be a control by which this goal can be achieved. This control could be the case of no control, f(t, x, u) = f(t, x) ; however, it must exist. Since it is not the purpose of these notes to delve into all the mathematical problems, it will be assumed that there exists at least one control by which the objective can be achieved. It will further be assumed that any control function used in the sequel yields a unique state function x (t) with x (t„) = x • the state func- tion is obtained by solving the dynamic equations. In general, there are a number of controls which could yield the desired system state. From among this set of possible controls, one would like to choose the "best" control with respect to some performance criterion. For example, one would like to choose the control so that the process is carried out with a minimum cost in fuel, or time, or money. In the sequel, it is assumed that the performance criterion can be expressed in terms of a cost function; furthermore, it is assumed that the cost function is additive with respect to the contribution from each time interval. An example of such a cost function is G(x T , T) + I F(a, x, u) da C where x = x(T) . This cost function is dependent on the final state of the system through the function G and on the intermediate states and the control function through the function F. The additive property of the control function with respect to the intermediate times is represented by the integral. By an optimal control is meant that control which minimizes the cost function; it is this function which is the desired result of optimal control theory. Any process that is being controlled is subject to unpredicted disturbances, and these can make a significant difference in the choice of a control function. Suppose the dynamic equations of a system is given by the differential equation f ="*'« where p(t) represents a disturbance. The behavior of the system in response to the two different controls (u 1 = - x) and (u„ = - e ) does not differ if there is no disturbance (p = 0) ; however, if a disturbance is present, the response is significantly different. If x_ = 1, the response to the first control is given by * f 3 t + e t e°p(c ) do J whereas the response to the second control is x = e + p(a) da ; Such differences could conceivably result in a different choice for an optimal control. In analyzing systems and their control, one must find a way to represent the unpredictable disturbances. Such disturbances cannot be modeled by analytic functions since the value of an analytic function at any point is predictable from its value on an arbitrary short interval. One answer to modeling these disturbances is to describe them as stochastic processes. The theory of such processes was developed to model the fluctuation observed in physical systems. Wiener processes or the Brownian motion process are of particular interest to the stochastic control problem; many of the disturbances that affect a control system can be modeled by processes generated from Wiener processes. A Wiener process is a stochastic process in which the statistical properties over the interval (t, t+x) are the same as those over the interval (s, s+t) ; moreover, the behavior of the process is independent over time intervals which do not overlap, and there is no trend in the behavior. Once the stochastic disturbances have been introduced into the control theory, the problem is no longer deterministic. The state variables and control variables are no longer predictable but must be 2 described by their statistical properties. Kalman and Bucy provide a solution to the stochastic control problem for nonstationary linear systems. Their solution consists of using an optimal filter to estimate from the observed system performance the state of the system in terms of the conditional mean; the estimated state is fed back to the control signal through linear feedback. The linear feedback is determined by solving a deterministic control problem; the filter depends on the disturbances and on the system dynamics, but it is independent of the cost. Although the nonlinear stochastic control problem or its equiva- lent, the nonlinear filter problem, has not been solved, some headway 3 has been made by Bucy and Joseph; this lecture considers only the linear problem. Astrom, K. J., "Introduction to Stochastic Control Theory," Academic Press, Inc., New York (1970). 2 Kalman, R. E. and R. S. Bucy, "New Results in Linear Filtering and Prediction Theory," Journal of Basic Engineering Series D, American Society of Mechanical Engineers, Vol. 83, pp. 95 — 108 (1961). 3 Bucy, R. S. and P. D. Joseph, "Filtering for Stochastic Processes with Applications to Guidance," Interscience Publishers, Inc., New York (1968). 4 As an example of a control problem, consider a ship moving through a current of water; the ship is a system undergoing a change in state. In this example, the state is the position (x, y) of the ship. The parameters which control the motion of the ship are the power, which determines the velocity relative to the water, and the steerage angle, which controls the heading angle 9. In this simplification of the system, the dynamic equations are: x = V cos + u(x, y) y = V sin 8 + v(x, y) where u and v are the velocity of the current in the x- and y-directions, respectively. The goal might be to go from point A to point B. If it is desired to reach B in the shortest possible time, the cost function would be the accumulated time; if it is desired to reach B with the minimum expenditure in fuel, the cost function would give the expended fuel in terms of x, y, V, and 0. A more complicated cost function would result if it is desired to reach B in the least time with a reasonable expenditure of fuel. Both the power and steerage angle could be subject to unpredictable perturbations; there could also be a stochastic pertur- bation of the current. This lecture on control theory first treats a deterministic optimal control problem with no constraints on the controls. It is first solved by transforming the problem into a boundary-value problem for an ordin- ary differential equation, the so called indirect approach; it is then solved by the direct method developed by Bellman, the method of dynamic 4 programming. The big contribution of modern control theory to the de- terministic control problem has been the extensions to controls with constraints, and a discussion of constrained controls constitutes an- other major topic of the lecture. Still another important area is the development of the theory of stochastic processes necessary in the 4 ,, „ Bellman, R. E. and S. E. Dreyfus, Applied Dynamic Programming, Princeton University Press, Princeton, N.J. (1962). treatment of stochastic controls. Finally, the theory of Kalman-Bucy filters is given and their solution to the stochastic control problem is presented for linear systems. THE OPTIMAL CONTROL PROBLEM In these lectures the simplest optimal control problem considered is that of a state variable x(t) and a control variable u(t) defined on an interval 0<t<T. The process being controlled is described by the dynamic equations x(t) = f(t, x, u) (1.1) with x(0) = x Q (1.2) The vector f_ is twice continuously dif f erentiable with respect to x and Lipschitz continuous with respect to u; this latter condition means simply that there is a constant L such that for every pair of control vectors u. and v |f_(t, x> u) - j[(t, x, v) | < L|u - v| (1.3) For each control vector u., these conditions imply that the state vector x> which is obtained from solving (1.1) and which also satisfies the initial condition (1.2), exists and is unique. Moreover, from among the set of control vectors, it is assumed that there is a unique control u which minimizes the cost function C . The cost function is defined by the following: T '0 C T [u] = G(x T , T) + J !■■(. , ,t) ,i '' '.) The functions F is twice continuously dif ferentiable with respect to x and Lipschitz continuous with respect to _u; G represents the cost at the terminal point x(T) = x_; it is twice continuously dif ferentiable with respect to X . Suppose that v is an optimal control vector, and consider a slight deviation 6_u of this control vector. If u(t) = v + 6u u_(t) is also a control vector, as can be seen from an application of the theory of ordinary differential equations. If z^ is the state vector associated with the control v, the new control _u yields a new state vector x given by x(t) = _z + 6x where 6x is an unknown. Moreover, since v minimizes the cost function, the new cost function is greater; j F(a, x, u) da + G(x ,T) J I F(o, z, v) da + G(z ,T) (1.5) Since the old state vector satisfies Z = f(t, z, v) and the new one satisfies x = f (t, x, u) then x + 6x^=x^=f^(t, z_ + Sx, v + 6u) Now by assumption f_ is twice continuously dif f erentiable with respect to 2c; hence 6x = f^(t, z^ + 6x, v + 6u) - jf (t, z_, v) = f 6x + f(t, z, v + 6u) - f_(t, z, v) + 0(|6^x| 2 ) (1.6) It is not necessary that 6_u be uniformly small; indeed, in problems involving bang-bang controls, this is not at all true. However, there can be deviations 6u of order one only if their duration is short. It can be proved that if 6_u satisfies the condition |ou(cr)| do < e (1.7) J then the deviations 6x(t.) are also of order e. Since by assumption, f_ is Lipschitz continuous with respect to u., \i_(t, z, u) - f(t, _z, v)| < L|u - v| = 0(6u) Moreover, it follows from Equation (1.6) that to the same order of approximation 6x = f 6 x + f_(t, z, u) - f (t, z^, v) (1.8) or in abbreviated form Sx = f 6x + f(u) - f(v) (1.9) — — x — — — This equation is a linear differential equation for 6x, and there are standard ways for solving linear differential equations. One first considers the linear homogeneous equation ± = Ay (1.10) where in our case y_ represents the vector 6x and A the matrix f . Let y j (t) = (SyCt, t), * 2j (t, T) $ nj (t, T)) be the solution of Equation (1.10) with $. . (x, x) = 6.., the Kronecker delta; moreover, let $(t, x) be the matrix whose column vectors are the vectors yj' , $(t, x) = $..(t, x) . The matrix $(t, x) is called the transport matrix or fundamental matrix for the differential Equation (1.10). From (1.10) it follows that as a function of t || (t, X) = A$(t, X) (1.11) and by its definition '(x, X) = I (1.12) where I is the unit matrix. The solution y(t) is given in terms of its value at t = X by y(t) = $(t, X) v_(T) (1.13) Coddington, E. A. and N. Levinson, "Theory of Ordinary Differential Equations," McGraw-Hill Book Company, Inc., New York (1955). Hence I Z(t) = y(t) = $(t, T) y(T) = $(t, T) $(t, t) y(t) or if J ^ 0, I = <I>(t, T) <J>(T, t) (1.14) Differentiating with respect to t yields = || (t, T) *(T, t) + $(t, T) |^ $(T, t) = A $(t, T) *(T, t) + $(t, T) |- *(T, t) at = A(t) + *(t, T) !^ $(T, t) It can be shown that $(t, t) has an inverse and that this inverse is $(t, t) ; consequently |- $(t, t) = - $ 1 (t, t) A(t) = - $(t, t) A(t) dt that is, $(t, t) as a function of T satisfies 4- $(t, T) = - *(t, T) A(t) (1.15) dT Although (1.15) will be used subsequently, of immediate interest is the solution to the inhomogeneous linear equation y = A y +£(t) (1.16) 10 with y(x) = 0; the solution is given by J «(t. y(t) = J $(t, o) g(a) da (1.17) T which can be verified by substitution into (1.16). For the control problem, (1.17) has two consequences: it can be used in conjunction with (1.6) to obtain an estimate for the order of magnitude of 6_x and it can be used to solve (1.9). In the first case, I 1 (i r(u) - l"(v) d j 0(6x 2 ) do i |*(t, a) I I f(u) - f(v)| do + | 0(6x' I | f (u) - f (v)|do + J 0(6x 2 ) do ,t < M where M is a bound for $. From (1.3) 5x| < LM J |Su| do + J 0(6x ) do ♦ r- < LMe + 0(6x ) do By iteration • r- : LMe : 0(e 2 ) do = 0(e) The second case is of more interest, of course, for it gives an approximation of 6x good to the second order in e, namely, 6x = *(t, o) [f(u) - f(v)] do (1.18) 11 where $ is defined by ff (t, T) = 4(t) *(t, T) (1.19) Now consider the difference in the values of the cost function; by (1.5) F(a, x, u) - F(a, z, v) da + G(x T) - G(z , T) > J Q T T Hence, from the assumptions on F and G, | [F 6x + F(u) - F(v)] da + G 6x(T) > ■A. X — — -x- — — i [F x (T > 1 By (1.18), ,T »(x, a) (f(u(a)) - f(v(a)))da + F( u (t)) - F(v(x))] dx + G j $(T, a) (f(u(a)) - f(v(a))) da > (1.20) x "b If the order of integration in the double integral is changed, .T a) [f(u(a)) - f(v(a)) dadx J F (x) J $(T, T T = J J F x (t)$(t, a) dx [f(u(a)) - f(v(a))] da (1.21) a T The vector function p_ is defined by „T F (x) $(t, t) dx - J x ^ (t) = - J F^(x) $(t, t) dx - G v (T) $(T, t) (1.22) t Recall that one of the properties of $ was (1.15) fr (x, t) = - $(t, t) f (t) 12 Then f p T = F (t) $(t, t) - | F (T) |r *(T, t — x J x ot ;t) + i F (T) *(T, t) f (t) dT + (t) - - J F (t) *(T, t) dx - G x (T) $(T, t) = F (t ) dx - G (T) ■§- *(T, t) x dt G (T)O(T, t) f (t) = F ■T T . £ = F „ " P I. f (t) (1.23) T T with p (T) = G (T). In terms of p , (1.20) becomes [- P (a) (f_(u) - f.(v)) + F(u) - F(v)] da > J [- F(v) + p T f(v)] - [- F(u) + p T f(u)] do > (1.24) Since 6u is an arbitrary deviation satisfying only (1.7), it can be chosen such that u = v everywhere except on some arbitrary interval; as a consequence, the inequality in (1.24) must hold for the integrand: - F(v) + p T f (v) > - F(u) + p T f(u) Define H(t, u) = - F(u) + p f (u) (1.25) Then H satisfies H(t, v) > H(t, u) (1.26) 13 for v, an optimal control. This is the Pontryagin maximal principle T which states that for given values of p_ and x at time t, the optimal control v(t) is the control function for which the Hamiltonian H(t, u) is a maximum. If the control functions are sufficiently smooth, the optimal control is that control for which = |S= - F + P T f (1.27) du u u It is assumed that f is dif ferentiable with respect to u_; prior to this equation, f need only be Lipschitz continuous with respect to u. This equation is a system of m equations which could be solved for the m control functions (u n ,...,u ) in terms of the state variables 1 m T T (x-,...,x ) and the new variables (p , . . . ,p ). Consequently, the optimal control problem has been reduced to a two-point, boundary-value problem for an ordinary differential equation: x = f(t, x, u) • T T7 T f p = F - p f = - F + p T f 9p T 9H 0=|» du (1.28) 14 where x(0) p(T) = G x (T) (1.29) T There are just enough conditions to determine x, p_ , and u. T The function H contains the variables x, £ , u, and of course t. Using (1.26) to eliminate u, (1.28) can be expressed in terms of the set T of dual variables x and £.'=£.> where the prime denotes transpose of the vector; the resulting system is the familiar canonical form of classical mechanics. 3H P = - I 1 (1.30) — dx The boundary conditions are stated in terms of x„, x. , and T; for instance, both x T and T might be fixed, or either one might vary while the other is fixed. No boundary conditions are specified directly in terms of p_; the boundary conditions on p_ are obtained indirectly by substitution into (1.29). Equation (1.29) does, however, contain a sufficient set of conditions to pose a two-point, boundary-value problem for (1.30). Another form that the boundary condition at t = T might assume is for x and T to satisfy an end condition of the form M(x T , T) = (1.31) where M is a twice continuously dif ferentiable vector function of both its arguments. In this case the method of Lagrange multipliers will be used to transform the optimal control problem into a corresponding 15 two-point, boundary-value problem. The vector q is introduced here as a Lagrange variable. Now the problem of minimizing the cost function (14) is replaced by the problem of finding the unconstrained minimum of f T C (u) = J F(a, x, u.) + £'(x - f_(a, x, u) da q + U'M(x T , T) + G(x T , T) (1.32) The boundary condition (1.31) has been inserted into the cost function by means of the Lagrange multiplier y_. Suppose v(t) is the control which minimizes C . For a variation 6u to the control v, let x(t) q - — — denote the new state variable, and let t = T+AT be the time at the new terminal point. The main difference from the previous argument in this section is that the terminal time is T+AT rather than AT. The new cost is given by T+AT C (u) = F(a, z^ + 6x, v + 6u) + £* (z + 6x - f (a, z + 6x, v + 6u) da q + G(x(T + AT), T + AT) + y'M(x(T + AT) , T + AT) Hence the increase in cost C (u) - C (v) is given as: q — q — ° C (u) - C (v) = [F 6x + F 6u + q'6x - q'f 6x - q'f 6u] da q - q + G x Ax T + G T AT + y ' (M x Ax T + M T AT) ,T+AT + J F(a, x, u) + £'x - £'f_(a, x, u) da T where x = z + 62 f T d6 ^ r T J q' -t— dt = q'(T) 6x(T) - q ' 6x da ~ dt ~ _ _ " 16 Hence ! r >i J {[F - q' - a'f ] 6x + [F - £'f ] 6u} da + q' (T)Sx(T) + G x Ax T + G^T + y' (M x Ax T + M^T) ♦ / T+AT F(a, x, u) + q/x - q_'f(a, x, u) - F(T, x u ) •T -T' - £'x(T) + q*f (T, x T , u T ) da + [F(T, x T , u T ) + £'z T + q'6x(T) - q'f(T, x T u,,) ] AT (1.33) The integral from T to T+AT is a second order contribution which goes to zero faster than the other terms as AT ->■ 0. In order to determine Ax , consider the solutions of the differ- ential Equation (1.1), which have the initial value x.. These solutions satisfy the integral equation Then :(t) = Xq + Ax T = (x(T) - z(T) r J !( a > E> u) dCT -AT ) + J f(a, x, H ) da = 6x(T) + f AT + 0(e ) where the geometry of the proof is illustrated in Figure 1. T T+AT Figure 1 — Geometry of the Proof 17 To within second order 6x(T) = Ax T - f AT (1.34) Within this order of approximation, (1.33) reduces to the following: T < I [F - q" - q'f ] 6x + [F - q'f ] 6u da — -L x — -x — L u — — u — + [q'(T) + G + y'M ] A x T — x — x — i + [-£' (T) f + G T + y'M T + F(T, x T> u. T ) ] AT If q is now determined so that the coefficient of 6x vanishes, q' = F - q'f — x — x This is the same differential Equation (1.23) that p_ satisfied; our Lagrange multiplier can then be identified with p_ q=£ (1.35) Moreover, since the relationship must hold independent of 6_u, F - p'f =0 (1.36) u u Since there are no longer restrictions on Ax and AT, p' (T) + u'M + G =0 — x x F + G T + y *M T - p'f = (1.37) Introducing the Hamiltonian (1.25) yields 18 9H 3p g d.28) o-|5 du The initial condition x(0) = x^ together with the terminal conditions M(x T , T) = p* (T) = - (y'M x + G x ) (1.38) H(T, u(T)) = G T + y'M T provides a sufficient number of conditions to determine x> £.» u, and T. The last two equations in the system (1.38) are obtained from (1.37). The problems of optimal control theory generally reduce to a two- point, boundary-value problem for the system of ordinary differential equations (1.30). Bailey, Shampine, and Waltman discuss methods for solving such two-point, boundary-value problems. These problems are presently solved either by the shooting method or by solving a sequence of simpler boundary value problems whose solutions converge to a solu- tion of the given problem. In any case, very few of these problems can be solved without the use of electronic computers either digital or hybrid. The shooting method is the easier, when it works. It consists of supplementing the conditions at one end with a sufficient number of assumed conditions to yield an initial value problem. The initial value Bailey, P. B. , et al., "Nonlinear Two-Point, Boundary-Value Problems,' Academic Press, Inc., New York (1968). 19 problem is solved; the solution is substituted into the boundary con- ditions at the other end. If these conditions are satisfied, the solu- tion to the initial value problem is the desired solution to the two- point, boundary-value problem; otherwise, a new set of assumptions is made based on the discrepancy between the actual boundary values and the calculated values. Hopefully, as one continues this iteration process, the solutions to the initial value problem converge to a solution of the two-point, boundary-value problem. The shooting method may not con- verge, or it can be unstable, that is, a small variation in the initial conditions results in a large variation in the solution. If the initial problem is unstable, a small error, such as roundoff on a computer, could cause subsequently computed values at another point to be meaningless. Before proceeding to the direct method for solving the optimal control problem, take a second look at the Hamiltonian H and the func- T tions p_ . Suppose that the terminal cost G is identically zero; the cost function is then T C(u) = ! F(a, x, u) da Further, assume that every point in an open neighborhood N of an optimal trajectory _z(t) can be joined to the initial point (0, jO °y a trajec- tory x(t) resulting from an optimal control. This assumption makes the minimal cost J a function of the terminal point (T, x T ) i- n N. J F(G, x, u) J(x , T) = Min J F(G, x, u) da (1.39) It is assumed that J is twice continuously dif ferentiable. Then, J(x T + Ax T , T+AT) = J(x T , T) + J x Ax T + J T AT (1.40) 20 By the definition of J, there is a control u + Su together with a trajectory x + &* such that J(x + Ax , T+AT) = J F(G, x + 6x, u + 5u) da (1.41) where u is the control such that T 1 J(x T , T) = J F(a, x, u) da (1.42) 1 From (1.41) and (1.42) J(x T + Ax T , T+AT) - J(x T , T) J F(a, x + 6x, u + 6u) - F(a, x, u) da .T+AT + J F(a, x + 6x, u + 6u) da T (F 6x + F 6u) dt + FAT x _ u " Now from (1.23), the equation f or j) is F = £' + £'4 Hence J(x T + Ax T , T+AT) - J(x T , T) f T = J [(£' + p'f ) 6x + F 6 H] da + FAT = p ' (T) 6x(T) + I £' (f 6x - 6x) + F 6u da + FAT x 21 From (1.9) Hence f 6x - 6x = - f 6u — x — — — u — J(x T + Ax T , T+AT) - J(x T , T) = p'(T) (Ax_ T - fAT) . r (F - p'f ) 6u da + FAT U ~^ ~ where use has been made of (1.34). But by (1.27), F p'f = 0; so, u u J(x T + Ax T , T+AT) - J(x T , T) = p'Ax T + (F - p * (T)f) AT = p'(T) Ax T - HAT = J x Ax T + J T AT where the last equality results from (1.40). This gives J x = P' (1.43) and J T = - H (1.44) In the space of variables (x , T) , the vector p_' is the gradient of the function J; it is normal to the surfaces of constant J; H is the Hamiltonian of the function J. This sheds new light on the maximal principle. Along an optimal trajectory, the change in cost J over a given time step AT is a minimum, that is, H is a maximum. 22 OPTIMAL TRAJECTORY J+AJ Figure 2 — Constant Cost Fronts These arguments hold only if the terminal cost is zero; G = 0. RELATION TO DYNAMIC PROGRAMMING The partial differential equations (1.43) and (1.44) can be ob- tained by the method of dynamic programming. This method is based on the Bellman principle of optimality. According to the Bellman prin- ciple, an optimal control policy has the property that, regardless of the initial state or initial decision, the remaining decisions must constitute an optimal control policy with regard to the state which results from the first decision. In terms of the cost function C(u) = J F(a, x, u) da the Bellman principle takes the form. The cost C(u) is a minimum along a curve x defined on [0, T] if it is a minimum along each later part of the curve, that is, if Dreyfus, S. E. , "Dynamic Programming and the Calculus of Variation, Academic Press, Inc., New York (1965). 23 J" T F(a, x, u) da is a minimum along the curve x. on the interval [t, T] for all te[0, T]. The integral is dependent on the end point (t, x(t)). If one defines r J(x, t) = min J F(a, x, u) da (2.1) u t for all admisible controls u, then I" 1 J(x, t) = min {F(t, x, u) 5t} + min J F da u u t+6t J(x, t) = min {F(t, x, u) 6t + J(x + 6x, t + 6t)} (2.2) This equation forms the basis of the direct methods for solving control 7 8 problems, described by Dreyfus. Larson extended the direct methods to constrained problems. If it is assumed that J has partial derivatives, the differential equations (1.43) and (1.44) can be obtained from (2.2). Hence, the boundary value problem for the optimal control is obtained. If the partial derivatives of J exist, the right-hand side of (2.2) can be expanded in a Taylor series: J(x, t) = min {F6t + J(x, t) + J (x, t) 5x + J (x, t) 6t} u (2.3) Larson, R. E., "State Increment Dynamic Programming," American Elsevier Publishing Company, Inc., New York (1968). 24 From the differential equation (1.1) Hence 5x = f6t = min {F + J f + J } 6t x— t Since St > 0, = min {F + J f + J } (2.4) x— t In order to find the minimum of the term in brackets, it is differ- entiated with respect to _u and the result is set equal to zero. This is a necessary, but not sufficient condition; however, if one assumes a minimum, it serves the purpose. + J fu = (2.5) x— By (2.4) Hence F + Jf+J =0 (2.6) x— t H = J =-H = -(F + Jf) t x (1.43) From (1.25) J = p T (1.42) 25 There is a difference between the definition of H here and its definition in the previous section. This is only an apparent difference in the sign of F, which occurs because the lower limit of the integral is used in the definition of J here rather than the upper limit as used earlier. Otherwise there is complete agreement with the results of the indirect method. CONSTRAINTS ON THE CONTROL AND STATE VARIABLES In most applications, the control or the state variables cannot be chosen arbitrarily but are subject to constraints. In the problem of a ship moving in a current, ship speed is limited by the maximum power available. The constraints can generally be expressed in terms of inequalities of the form £(x, u) < (3.1) where the vector inequality simply means that the components satisfy the inequality. The number of components in the vector _£ is the number of constraints on the system. The analysis does not depend on whether both x and _u occur implicitly in the inequality; one can have constraints on the controls and not on the state of the system or vice versa without affecting the analysis. In this presentation, the variables in the optimal control problem with constraints are the state variable x(t) and the control variable _u(t) defined on an interval <^ t £ T. The process being controlled is described by the dynamic equation (1.1): x(t) = f(t, x, u) with initial condition x = x ; the state and control variables are — — c constrained by the inequality (3.1). For simplicity, the terminal cost is taken as zero, G = 0, and the cost function is given by the equation: ,T ,(u) = J C (u) = J F(a, x, u) da (3.2) 26 The vector _f and the cost function F are twice continuously dif ferentiable with respect to x and continuously dif ferentiable with respect to u. The Lagrange multipliers will be used here to reduce this problem to a two-point, boundary-value problem. As in (1.32), the differ- ential equation is introduced into the cost function by means of a Lagrange multiplier p_. f T C (u) = F(o, x, u) + p'(x - f(a, x, u)) da p - J Q which yields the variational equation f 1 ) [F 6 +F6u+p'6x-p'f6x-p'f 6u] da J x— x u— — — — ^x — — — u — . r p_'(T)6x(T) + I [(F - £' - p'f ) 6x + (F - p'f ) 6u da > (3.3) u u — — The differential in the cost is greater than or equal to zero since it is assumed that the variation 6u is around an optimal control, a control which minimizes the cost. Because of the constraint (3.1), the vector 6u is not free. For instance, suppose that for t between \. and t„, the trajectory z_(t) due to the optimal control v(t) is along the boundary of the allow- able region; see Figure 3. One cannot freely choose the variation 6u_ in the control vector for t < t _< t„ and still expect to remain in the allowable region R. For the optimal trajectory z^ and control v, there are at most a finite number of intervals t < t < t + 1 such that equality holds for any of the equations in (3.1).* On such an interval, the conditions (3.1) can be split into two sets *The proof of the statement is topological and beyond the scope of these notes. 27 z(t) x(t) = Z + Figure 3 — Constrained Variables and ^(z, v) = ,(z, v) < (3.4) where ^ = ($>_ £ ) . Consider a new vector ^ defined by ^(x, u) + _^(x, u) = The vector ip is called a defect vector. Along the optimal trajectory, the vector ^_ can also be split into two component vectors, _^ and _^„ , which correspond to the component vectors of $>_. The component vectors of \\i also change from interval to interval. Along a given interval [t k> C k + 1 ] 28 4 = o (3.5) ± 2 > ° Since ^ (z^ v) is zero on this interval, either jz, v, or both are on the boundary of their allowable range. From previous arguments, it is known that one cannot freely choose 6u_. Only those values of &u are allowed which satisfy A (z + fix, v + 6u) £ or by (3.4) 4Li (z + fix, v + 5u) - j>_ (z, v) £ On the other hand, for a neighboring trajectory to _z A ( z^ + fix, v + 5u + _^ + _6J) = on [t k' hc+i^ since i = - i <j>(z + fix, v + 6u) - <J>(z_, v) + 6^ = (3.6) In order that the above inequality and (3.6) hold, 6^ > (3.7) Moreover, provided the variations are sufficiently small, 6^„ is free. If _<£ is twice continuously dif f erentiable, then it follows from (3.6) that <j> fix + cj> 6u + &\l) = (3.8) 29 Set Su = (611-. , 6u„) and consider the first N, equations in (3.8), — — i — I (p N^ = dim (^). If the square matrix (J> is not singular, its inverse y exists, and 6u x = - Y i lu2 6u 2 - Y 4 lx 6x - y 6^ The vectors 6_u and 6x are free; the vector 6_^ satisfies (3.7). If the matrix <t> n is singular, the first N, constraints were de- 1 *1 pendent; eliminate the dependent constraints and start again. The contribution to the cost differential (3.3) from the interval t, < t < t, is the following integral: r fc k+l \ - J { ^ ~ P-' " £'4 " V Y i lx " £'f Y 4x ] 6 ^ fc k X X + [(F Uz " P'^ - F^ - P'^) Y4 1U2 ] ^ - (F -p'f ) y 61 } da u J u 1 -"-1 Define the vector X by X' = - (F - p'f ) Y ( 3 - 10 ) -1 u x -^ Then \ = C k + 1 C k {[F x -p- - P'f x + Ali lx ] + r^-p'i^+AIi^] 6u 2 + xi6V da 30 The vector £ can be determined so that the coefficient of 6x vanishes: p' = F - p' f + X'j. (3.11) — —x — ~~x — 1— lx Since 6u„ is free, the usual argument that Su is zero everywhere except on a small interval yields F - p' f + X' 4, =0 (3.12) u 2 * -u 2 -1 -lu 2 Now 6u can be chosen so that 6u = for t < t, and for t, , , < t. — — — k k+1 — In this case, the only contribution to the cost difference (3.3) is that due to I, ; hence k < I, = A' 6^i da By (3.7), ty > 0; so A > (3.13) Let the Hamiltonian be defined by H=-F+p' f - X' £ (3.14) where A is defined by A. > if *. = A. = if 6. < 31 The differential system (1.28) also holds for this H, that is, 3H p=-|? (1.28) 8x du One example of a constrained control problem is that of a forced harmonic oscillator in which the magnitude of the force is limited. In this problem, the force is the control and the process is one of chang- ing the velocity and displacement of the harmonic oscillator. It be- comes an optimal control problem if one is interested in finding the force or control which reduces the oscillator from a given velocity and displacement to zero velocity and displacement in minimum time. The equation of motion for the forced harmonic oscillator with a limited force is simply ^f + cz = F dt where |f| _< M, a given constant. Set x = cz/M, x = tot, and u = F/M where co = /c/M. In terms of these nondimensional variables, the non- dimensional form of the equation of motion is x + x = u (3.15) where the control function satisfies the inequality |u| _< 1. This constraint can also be written in the form 32 ^(u) = (u - 1) < (3.16) (u) = - ( u + 1) < The optimal control problem can be formulated in the phase plane. If (x, y) are the phase plane coordinates, the equation of motion (3.15) takes the form (3.17) y = - x + u Starting the oscillator at a given displacement with a given velocity is equivalent to assigning a given point (x, y) = (a, b) in the phase plane as an initial condition for (3.17). The rest state of the oscillator is represented in the phase plane by the point (0, 0), the point of zero displacement and velocity. Hence, the optimal time control problem is one of finding a control u which minimizes the time between states (a, b) and (0, 0). In this problem, the cost is given by -f C T (u) = T = J dT (3.18) 1 The cost function F(x, x, u) = 1. Set p = (p, q). Then the Hamiltonian defined by (3.14) is H = - 1 + py + q(u - x) - X(u - 1) (u + 1) (3.19) and, moreover, (1.28) takes the form 33 3H X = 8? = y 9H o. = - x + u 3q = q 9H 8x 3H = |^ = q - A(u - 1) - A(u + 1) (3.20) Suppose u is an optimal control which reduces the oscillator from the state (a, b) to the state (0, 0) in the minimal time T, and suppose |u| < 1 for the interval T < T < T . Suppose q + on T < T < T . By (3.20), q - 2Au = 0; hence, A i on (T , T ). A consequence of A i is that (f) = 0; hence, if q ± 0, it follows that |u(t)| = 1 on (T , T ) In other words, one needs to look only for the optimal control among those controls for which |u(t)| = 1. Now |u| =1 implies u = + 1; hence, the solution of (3.20) is given as: x + 1 = A sin (t + a) y = A cos (t + a) p = B sin (t + a ) q = B cos ( T + a„) q = 2 A u (3.21) 34 Since A > 0, it follows from the last of these equations that the sign of q is the same as the sign of u. Hence, if q changes from positive to negative, the optimal control must switch from +1 to -1. It switches from -1 to +1 if q changes from negative to positive. In a neighborhood of the origin, the optimal trajectory satisfies 2 2 (x + 1) + y = 1 Hence, its final segment is either on the circle of radius 1 about (-1, 0), or it is on the circle of radius 1 about (1, 0); see Figure 4. Suppose for the sake of argument that there is an £ > such that u(t) = - 1 for T - £ _< x < T. The last segment of the optimal trajector 2 2 is on the semicircle { (x + 1) + y = 1, < y}.. Between (0, 0) and (-2, 0), the parameter T would change along this semicircle by the amount tt; hence, the sign of q must change somewhere on this semicircle. At the point S 1 where q changes sign, the sign of u must also change, and u switches from -1 to 1. The optimal path continues backward on the circle of radius r around (1, 0) until either (a, b) is reached or q changes sign. But q does not change sign until the point S„ is reached since the time between S and S is tt. At S , the control would switch to -1 and the optimal trajectory would continue back on the circle of radius r„ around (-1, 0). This process is continued until the point (a, b) is reached. In the process, one switches control each time one of the following semicircles is intercepted: [x - (2n - l)] 2 + y 2 = 1, y > 0, n=0,l,2,... (3.22) [x + (2n - l)] 2 + y 2 = 1, y < n=0,l,2,... (3.23) The curve formed by these semicircles is called the switching curve; see Figure 5. 35 (a, b) Figure 4 — Optimal Trajectory Figure 5 — Switching Curve 36 The optimal control and the resulting trajectory in the phase plane can now be obtained by reversing the above procedure. If (a, b) is above the switching curve, proceed with the control u = - 1. The optimal trajectory will be along the circle 2 2 2 2 (x + 1) + y = (a + 1) + b in the direction of that part of the switching curve which lies to the right of x = 0. For (a, b) on the switching curve, use u = - 1 if x < or u = 1 if x > 0. If (a, b) lies below the switching curve, start with u = 1 and change to u = - 1 at the switching curve. Change the sign of u at each intersection with the switching curve. When u = 1, the optimal trajectory lies on a circle with center at (1, 0); when u = - 1, it is on a circle around (-1, 0). Suppose only one switch in u is needed to reach the origin from (a, b) . Because of the symmetry of the problem geometry in the phase plane, it is necessary to consider only those cases for which a = 1 after the switch. The origin is then approached along the trajectory x = 1 - cos (T - t) y = - sin (T - t) (3.24) 2 2 which is on the semicircle {(x, y) | (x - 1) + y = 1, y <_ 0} let T be the time at which the switch occurs. The optimal trajectory s for T < T is given by — s x = - 1 + A sin (t + a) y = A cos (T + a) (3.25) 37 where A and a are constants defined by A sin a = a + 1 A cos a By (3.24) and (3.25), the switching time must satisfy 1 - cos (T - T ) = 1 + A sin (x + a) s s - sin (T - x ) = A cos (x + a) s s Elimination of T from these equations yields a relationship between the terminal time T and the initial point (a, b) , namely, (a + 1 + cos T) 2 + (b + sin T) 2 = 4 (3.26) By definition, time fronts are the curves which connect initial points having the same terminal time T. Equation (3.26) can be used to determine the time fronts for T <_ u . If T = 0, the time front is simply the origin; if there are no switches in the control, the initial point is an endpoint of the curve connecting all initial points from which the origin is reached with one switch in time T. More than one switch would require T > ir. From (3.26), the time fronts for < T < it are segments of the circle of radius 2 around the point (-1 - cos T, - sin T) ; see Figure 6. It is the segment of the circle which lies above the switch- ing path. At the switching path, the time front is tangent to the vertical line x = constant for x > 0; at the opposite end, it is tangent to the switching curve. For T = tt, the time front is a circle of radius 2 around the origin. 38 Figure 6 — Time Fronts 39 STOCHASTIC SYSTEMS Stochastic control theory was first applied in this country at the Massachusetts Institute of Technology during World War II to synthesize fire control systems. In the 1960's it was applied to space navigation, guidance, and orbit determination in such well-known missions as Ranger, Mariner, and Apollo. Applications of the filtering theory, aspects of control theory include submarine navigation, fire control, aircraft navigation, practical schemes for detection theory, and numerical in- tegration. There have also been industrial applications; one example involved the problem of basic weight control in the manufacture of 1 paper. The filtering and prediction theory developed by Wiener and Kolmogorov forms the cornerstone of stochastic control theory. It provides an estimate of the signal or the state of a process on the basis of observa- tion of the signal additively corrupted by noise. Unfortunately, the Wiener-Kolmogorov theory cannot be applied extensively because it requires the solution of the Wiener-Hopf integral equation. It is difficult to obtain closed form solutions to this equation, and it is not an easy equation to solve numerically. 2 Kalman and Bucy give a solution to the filtering problem under weaker assumptions than those of the original Wiener problem. Their solution makes it possible to solve prediction and filtering problems recursively and is ideally suited for digital computers. Basically, it can be viewed as an algorithm which, given the observation process, sequentially computes in real time the conditional distribution of the signal process. The estimated state of the process is given as the output of a linear dynamical system driven by the observations. One determines the coefficients for the dynamical system by solving an initial value problem for a differential equation. This differential equation is easier to solve than the Wiener-Hopf equation. Our attention here will be limited to linear systems with quadratic cost functions. In this case the solution of the optimal control 40 problem is given by the separation theorem. The solution consists of an optimal filter for estimating the state of the system from the ob- served data and a linear feedback of the estimated state of the system; see Figure 7. PROCESS OBSERVED DATA ^CONTROL SIGNAL OPTIMAL FILTER LINEAR FEEDBACK ESTIMATED STATE Figure 7 — Stochastic Control System The optimal filter is the Kalman-Bucy filter, which will be dis- cussed in detail in the next section; the linear feedback is the same as would be obtained if the state of the system could be measured exactly and if there were no randum disturbances in the system. Thus, the linear feedback can be determined by solving a deterministic problem. Because of time limitations, we will not prove but merely accept the separation theorem. One objection to the use of stochastic control theory is that the process to which the theory is applied may not be random but merely irregular. For instance, the traffic flow on the Washington Beltway may not be truely random but it is certainly highly irregular. If I need to reach Dullis Airport from DTNSRDC by 1 pm, it might take me 45 to 50 minutes; but to reach the airport at 6 pm, I would have to allow 2 hours. The reason for this variation in lead time is that there will be bumper-to-bumper traffic on the Beltway during the rush hour and any accident brings this traffic to a halt. It is not the microscopic but the macroscopic properties of the traffic flow that govern our lead time estimate. The traffic flow could be analyzed as a stochastic process; such a model would be acceptable provided it predicted the macroscopic 41 properties of the flow. This is analogus to using linear models in the deterministic case. If the predictions agree with the experimental results, the linear theory is said to be good; if they do not, then the process is said to be nonlinear. In using a statistical model, one should recognize that it is only a model and not the actual process, and one should continually strive to determine the accuracy of his models. There are many reasons in favor of applying stochastic theory. The solution of the stochastic problem may be possible whereas the determin- istic theory may be hopelessly impossible. In many problems such as that of traffic flow, one may not be interested in the microscopic properties but merely in certain macroscopic properties. In the control problem, the stochastic model distinguishes between open and closed looped systems but the deterministic model does not. Another reason for using a stochastic model may be that this model is closer to the physics of the actual situation. In any case the purpose of this section is to lay the ground work for stochastic control theory. Our attention will be focused on certain concepts of stochastic processes and random differential equations. To describe a stochastic process rigorously would require measure theory and a great deal more time. Our approach will therefore not be rigorous, but hopefully it will be complete enough to get across the 9 basic ideas. For the rigorous approach, see either Doob or Gikhman and Skorokhod. A real random variable E, is a set of numbers or events together with a probability measure defined on this set. It is characterized by its distribution function F(x) which is defined by F(x) = P {E, < x} 9 Doob, J. L. , "Stochastic Processes," Wiley, Inc., New York (1963). Gikhman, I. I. and A. V. Skorokhod, "Introduction to the Theory of Random Processes," W. B. Saunders Company, Philadelphia, Pa. (1969). 42 where P {E, <_ x} is the probability that E, is less than or equal to x. The distribution function is nonnegative, nondecreasing, and continuous from the left; also F(- °°) = and F(°°) = 1. Analogously, if E, is an n-truple of random variables, its distri- bution function is a function of n real variables. F(x r x 2 ,..., x n ) =P {q<x r ..., ?n <x n } and F is called a joint distribution function of the variables E, . The function F(x.. , x„,..., x ) is uniquely defined in n-dimensional Euclidian space E , is non-decreasing, and is continuous from the left with respect n r to each variable. Furthermore, F(x r x 2 ,...x., - oc, x . +2 ,...x n ) = and F(x r ..., x., °°,..., oo) = F (i) (x 1 ,..., X.) where F denotes the distribution function of the i-truple (E, , . . . , E,.). A random function or a stochastic process is a random variable E,(t) which is a function of time. As time varies, £(t) describes the evolu- tion of the process. If a random process is recorded as it evolves, the recorded function £(•) describes only one of the many possible ways in which the process might have developed. The recorded function £(*) is called a sample function of the random process. For each fixed value of t, the quantity £j(t) is a random variable. Whereas a random variable is characterized by a distribution function, a stochastic process is characterized by a set of joint distribution functions. Assume that it is possible to assign a probability distribution to the multidimensional random variable E, = (E,(t ), £(t„),..., £(t )) for any n and arbitrary times t.. The distribution function 43 F(x r x 2 . . . . , x n ; h t n ) = P (C( tl ) < x x S(t n ) < x n } is called the finite-dimensional distribution of the stochastic process £(t). For F to be a distribution, it must satisfy the following com- patibility conditions: F(x r x 2 ,..., x., oo,..., oo; tr ..., t n ) = F (x ls x 2 ,..., x.; ^ t n ) for i < n and F(x,,..., x ; t 15 ..., t ) = F(x. ,..., x. ; t. ,..., t. ) 1 n 1 n j ' 3 3-, jn J l J n J l J where j .,,..., i is an arbitrary permutation of the indicies 1, 2,..., J l J n J r ' ' ' The mean value of a stochastic process is defined by oo m(t) = E[£(t)] = J x d F(x, t) where E is the mathematical expected value. The mean value is thus a function of time. Higher moments of £ are defined similarly. The covariance of the stochastic process is given by r(s, t) = cov [£(t), £(s)] = E [(£(t) - m(t)) (£(s) - m(s))] (x - m(t)) (y - m(s)) d F(x, y; t, s) oo Our definition of a stochastic process is very general, and most systems which come under this definition would be mathematically unmanageable. Some specialization of the theory which makes it possible to characterize the distribution of ^(t ), £(t„),..., £(t ) in a simple way are particularly attractive. For instance, if the distribution of 44 £(t, ) , . . . ,E,(t ) is identical to the distribution of E,(t + t) , 5(t„ + x),...,£(t + t) for all x and all arbitrary choices of the times t, ,....t , then the stochastic process E(t) is said to be stationary. 1' n r If only the first and second moments E[£] and E[^ ] of the distributions are equal, then the process is weakly stationary. Our discussion of control systems has been limited to systems in which knowledge of the system at time t together with the governing equations suffices to describe its future evolution. Knowledge of the past when the present is given is superfluous relative to the future evolution of the system. The stochastic system analogy of this situation is the Markov property for random processes; these are stochastic process- es in which the past and future of the processes are conditionally independent. In order to define a Markov process, the conditional probability and the transition probabilities have to be defined. The conditional probability P(a|b) is the probability that A will occur if B has occurred. Given a sequence of times t < t„ <...< t < t, the probability that £(t) < x if the sample function £(•) has already taken the values £(0, £(t_) , . . . ,£(t ) is denoted by P(£(t) < x|£(t ),..., L Z n — 1 £(t )). A stochastic process is said to be a Markov process if n p(£(t) < x|^(t 1 ),..., C(t n )) = P(£(t) < x|?(t n )) The transition probability distribution F(x, t|y, s) is defined by F(x, t|y, s) = P(£(t) < x|C(s) = y) . If a stochastic process is a Markov process, its finite distribution functions are given by F(x r x 2 ,..., x n ; t r ..., t n ) = F( Xl ; t x ) F(x 2 , t 2 | Xl , tl )...F(x n , tjx n _ r t^) 45 This results from an application of the Baye rule. A Markov process is thus defined by two functions, the absolute probability distribution F(x, t) and the transition probabilities F(x, t|y, s) . Consider a system with the following dynamic equation: x = f(t, x, u) + £ w (t) (4.1) where ^ is a small parameter and w is a stochastic process. Since w is stochastic, the state of the system x will also be stochastic; thus, we are interested in solving stochastic differential equations. Further- more, our interest is not with a particular sample function x(*) which is a particular discription of the state of the system during one run through the process; our interest is with the statistical properties of the stochastic process x (t). Consider the linear stochastic differential equation dx = A x dt + dw (4.2) where w is a stochastic process. In order to make some progress in finding the statistical properties of x, assume that w is a Wiener process . A Wiener process is a Markov process which satisfies the following conditions: 1. It is a second order process; that is, for all t E[w (t)] < <*> Hence, the mean m(t) exists as well as the covariance function r(s, t) = cov [w(t), w(s)] 46 2. The process has independent increments; that is, for arbitrary times t, < t„ < . . . < t , the increments 12 n x(t n ) - x(t n _ 1 ), x(t n _ 1 ) - x(t n _ 2 ),...,x(t 2 ) - x(t 1 ), x(t 1 ) are independent.* 3. The distribution of x(t) - x(s) for arbitrary t and s depends only on t-s. In this case, the process is said to have stationary increments. 4. The transition probabilities are Gaussian. In the one- dimensional case, the transition probability density is p(t + At, w|t, 0) = ^^ exp - w /2At 1_ 2 \t 5. w(0) = with probability one, and E[w(t)] = for all t > 0. Sample functions of a Wiener process have interesting properties. They can be continuous functions but are nowhere dif ferentiable. Their paths are of infinite length. Yet it is for just such perturbations that (4.2) will be solved. If w in (4.2) had bounded variation, the solution could be written in terms of the transport matrix <J>(x, t) of the linear system y = A y (4.3) The solution of (4.2) would be x(t) = $(t, 0) c + J $(t, t) d w(x) (4.4) where the value of x at t = is the random variable c. The expectation of c is m and its covariance matrix is T. 9 ^Independent random variables are defined on page 7 of Doob. 47 The integral f J $(t, t) d w(t) is a stochastic integral. Since the transport matrix <J>(t, t) is deterministic and has continuous derivatives, one way of defining this integral is through integration by parts. J <D(t, t) d w(t) = $(t, t) w(t) - $(t, 0) w(0) - J ||- (t, T) w(t) dT J o 9t It follows from (1.15) and other properties of the transport matrix that J $(t, t) d w(x) = w(t) - $(t, 0) w(0) + ! 4(t, T) A(t) w(t) dx (4.5) The integral on the right exist for almost all sample functions since the sample functions of w(t) are almost all continuous. This way of defining the integral has the desirable feature that the integral can be interpreted as an integral of sample functions. It does not, how- ever, preserve the intuitive idea that the integral is a limit of sums of independent random variables nor can it be extended to the case where $ is stochastic. Doob gives a more formal definition of the integral together with detailed proofs of its stochastic properties. The expected value of this integral is computed as follows: 48 I J $(t, x) d w(t; = E[w(t)] - $(t, 0) E[w(0)] + E $(t, T) A(T) w(T) L (3 dT = m(t) - $(t, 0) m(0) + j $(t, t) A(t) m(x) dx Hence $(t, x) d w(t) = J ${t, x) d m(x) L o Jo (4.6) The properties of the solution of the stochastic differential equation (4.4) will now be investigated. Since x is a linear function of a normal process, it is also normal and can be characterized com- pletely by the mean value function and the covariance function. Since the expected value of the Wiener process w(t) is zero, E[x(t)] = $(t, 0) E[c] + E = $(t, 0) m Q J $(t, T) d w(x; where m„ is the expected value of the initial condition c. Hence (t) = E[x(t)] = $(t, 0) m (4.7) Taking derivatives yields dm TT = 77 *(t, 0) m n = A(t) <D(t, 0) m n = A(t) m (4.3) at at U u x 49 Thus the mean value satisfies the linear differential Equation (4.3). The covariance matrix is more difficult to compute. In order to simplify the calculations, assume m_ = 0; hence, E[x(t)] = 0. This can always be achieved by subtracting m from x. For s >_ t, R(s, t) = cov [x(s), x(t)] = E[x(s) x (t)] "! r s It" j $(s, t) x(t) + J *(s, a) d w(a) 1 x (t) = E = $(s, t) E[x(t) x (t)] + J 4(s, a) E[d w(a) x (t)] t = $(s, t) R(t, t) (4.9) The integral is zero since w(a) and x(t) are independent for s _> t. T Set P(t) = R(t, t) = E[x(t) x (t)]. Then P(t) is the variance and is therefore the function of interest. P(t) = E '(t, 0) c + J $(t, T) d w(x)j ( $(t, 0) c + J $(t, a) d w(a) j = *(t, 0) E[c c T ] $ T (t, 0) + $(t, 0) E I c I d w (a) <£> (t, a) . / >(t, t) E[d w(t) c (t, 0) + J J $(t, t) E[d w(t) d w T (a)] $ T (t, a) 50 The increments of the Wiener process are independent of C; hence E [c d T w(a)] = E [d w(x) c T ] = Moreover, from the properties of the Wiener process E [d w(T) d w T (a)] = if dx and do have no parts in common; otherwise E [d w(T) d w T (x)] = R dT w where R is the covariance matrix of the Wiener process w. The final w expression for P is then P(t) = $(t, 0) T 4> T (t, 0) + $(t, x) R (t) $ T (t, x) dx J (4.10) A differential equation for P can be obtained from this expression for P simply by differentiating dt " [ dt $(t ' 0) 1 r $T(t ' 0) + $(t ' 0) r dF $T(t> 0) + $(t, t) r (t) $ T (t, t) + ; 9 $ ^' T) R (x) $ T (t, t) w «L dt w c + $(r, t) R ("i J „ w dT (t) ^r $ T (t, T) dT at 51 The transport matrix satisfies ^ (t, t) = A $(t, T) and Hence 3t 9 * £' T) = $ T (t, T) A T 4^ = a $(t, o) r $ T (t, o) + $(t, o) r $ T (t, o) a t at + R (t) + ! A $(t, T) R (t) $ T (t, + $(t, T) R (T) $ (t, T J„ w T ) A dT 4^ = A I $(t, 0) r $ T (t, 0) + $(t, T) R (T) $ T (t, T) dT dt ( • J Q w + \ »(t, 0) r $ T (t, 0) + J $(t, T) R $ T (t, t) dT [ A 1 \ ~ I + R (t) Thus from (4.10) dP T ^ = A P + P A + R //nN dt w (4.11) P(0) = T (4.12) 52 THE KALMAN-BUCY FILTER The solution of the optimal control problem for a linear stochastic system is given by the separation theorem. It consists of an optimal filter for estimating the state of the system from the observed data and a linear feedback of the estimated state of the system; see Figure 7. The linear feedback is the same as the feedback that would be obtained if there were no stochastic perturbation of the system. This section will develop the explicit computational schemes for solving the filter- ing problem. Suppose we have the stochastic process described in the previous section dx = A x dt + d w(t) (5.1) x(0) = c (5.2) where w(t) is a Wiener process and c is a Gaussion zero mean n-vector. In an actual case in which the process is realized, it is important to know the state of the system. It is, however, not always possible to measure x directly; instead, a set of quantities z(t) dependent on x are measured. Assume that the dependence of z on x is linear and is given by dz = H x dt + dv (5.3) where the perturbation v is a Wiener process independent of x. The filter problem can be formulated as follows. Assume that a realization of the output z has been observed over the interval < T < t. Determine the best estimate of the value of the state vector x at time t. It is assumed here that the admissible estimates of x are linear functionals F(z) of the observed output z. The criterion 53 for determining the best estimate is that the mean square estimation error be a minimum. This best estimate x(t) is dependent on the values of z(t) in the interval < x < t, and it can be proved that it is a linear combination of the values of z on this interval. ;t) = J K(t, t) d z(t; x(t) = J K(t, T) d z(T) (5.4) Since z(l) is a stochastic variable, x(t) is a stochastic integral. Interpolation and extrapolation are two problems that are related to the filtering problem. The interpolation problem is one of estimating the state at some time T < t; the extrapolation problem is one of esti- mating it at some time I > t. This latter problem is the one which is of interest to the stock market investor. The condition that x(t) is the best estimate from among all linear functionals of z(t) for the state vector x in the least squares sense is stated mathematically as follows. For every constant vector A. and linear functional F, E[{A T (x(t) - x(t))} 2 ] < E[U T (x(t) - F(z))} 2 ] (5.5) where all variables have a zero mean. E[x(t)] = E[x(t)] = E[F(z)] = Now set x = x - x where x is called the minimum error vector. E[(A T x) 2 ] < E[A T (x + (F(z) - x)) 2 ] < E[(A T x) 2 ] + 2E[A T x A T (F(z) - x) ] + E[(A T (F(z) - x)) 2 ] 54 For all A and F(z), the criterion (5.5) requires E[(A T (F(z) - x)) 2 ] + 2E[A T x A T (F(z) - x) ] > This can be true only if = E[A T x A T (F(z) - x)] = A T E[x(F(z) - x) T ]A But this implies that E[x (F(z) - x) 1 ] = for any linear combination F(z) of elements of z; hence E[x F (z)] = (5.6) An integral equation for the kernel K(t, t) can be derived from (5.6). This kernel is not a stochastic quantity, and it can be de- termined independent of the realization z(*)- For F(z) = z(x) - z(a), < a < x < t, the expression (5.6) yields E[x(t) (z(t) - z(a)) T ] = E[x(t) (z(t) - z(a)) T ] = E = E = E = E x( «(f" (s) x(s) ds + d v(s) K(t, r) dz(r) [ \ H(s) x(s) ds + dv(s) K(t, r) (H(r) x(r) dr + dv(r)) (H(s) x(s) ds + dv(s)) J K(t, r) H(r) x(r) x T (s) H T (s) ds dr a .t „T a a 55 I K(t, r) H(r) x(r) dv (s) dr J o + J I K(t, r) dv(r) x a + J I K(t, r) dv(r) dv T (s) T T (s) H (s) ds a From the properties of Wiener processes, .t „T -T E J | K(t, r) dv(r) dv T (s) = J K(t, s) R (s) ds a a where R is the covariance matrix of the process. Furthermore, v T dv(s) and x(s) are independent, so E dv(r) x (s) = 0; hence E[x(t) (z(T) - z(0)) T ] = J j | K(t, r) H(r) E[x(r) x T (s)] a I (5.7) H (s) dr + K(t, s) R v (s) } ds for all a and T. On the other hand, from (5.3) ;o If h E[x(t) (z(T) - z(0)) ] = E x(t ^ T (s) x(s) ds + dv(s) O J E[x(t) x T (s)] H T (s) ds J j J K(t, r) H(r) E[x(r) x T (s)] H T (s) dr a \ + K(t, s) R (s) } ds 56 where the last equality results from (5.7). Since this equation holds for all a and T in the interval [0, t], K(t, s) R (s) = E[x(t) x T (s)] H T (s) - ! ; K(t, r) H(r) E[x(r) x T (s)] H T (s) dr (5.8) This is a nonhomogeneous integral equation for K(t, s) . Its kernel T T is H(r) E[x(r) x (s)] H (s) . Since it corresponds to a positive definite quadratic form, all its eigenvalues are positive and the equation has a solution. Unfortunately, it is not possible to calculate K(t, s) from T this equation because E[x(r) x (s) ] , the covariance of x(s), is unknown. A different equation for K(t, s) can be obtained from (5.8) by differentiating both sides of it with respect to t. 3 K( ^ S) R (s) = f- E[x(t) x T (s)] H T (s) at v at - K(t, t) H(t) E[x(t) x T (s)] H T (s) - J 3 K( *» r) H(r) E[x(r) x T (s)] H T (s) dr By (5.1) Hence dx = Ax dt + dw E[dx(t) x T (s)] H T (s) = A(t) E[x(t) x T (s)] H T (s) dt + E[dw(t) x T (s)] 57 The second term vanishes since dw(t) and x (s) are independent if t 2l s. This yields S U ll S) R (s) = [A(t) - K(t, t) H(t)] E[x(t) x T (s)] H T (s) OL V -J || (t, r) H(r) E[x(r) x T (s)] H T (s) dr Use of the integral equation (5.8) to obtain an expression for E[x(t) x T (s)] H T (s) yields 9 K(t, s) + K(t, t) H(t) K(t, s) - A(t) K(t, s) R (s) ~r K( l:' r) + K(t, t) H(t) K(t, r) - A(t) K(t, r) at H(r) E[x(r) x T (s)] H T (s) dr (5.9) Set iKt, s) = 9 K( ^' S) + K(t, t) H(t) K(t, s) - A(t) K(t, s) — ot Then by (5.9) H t, s) = - J ±(t, r) H(r) E[x(r) x T (s)] H T (s) R^ 1 (s) dr (5.10) Since the kernel of this integral equation corresponds to a positive definite quadratic form, the only solution of (5.10) is ?(t, s) = 58 This yields the following differential equation for K(t, s) 9 K(t, s) 3t = A(t) K(t, s) - K(t, t) H(t) K(t, s) (5.11) From the integral equation (5.8) for K(t, s) K(t, t) R (t) = E[x(t) x T (t)] H T (t) On the other hand, T. E[x(t) x (t)] = E J K(t, r) H(r) E[x(r) x T (t)] H T (t) J K(t, r) dz(r) x T (t) = J K(t, r) H(r) + J K(t, dr E[x(r) x (t)] dr r) E[dv(r) x (t) ] where the second integral vanishes; hence K(t, t) R v (t) = E[x(t) x T (t)] H T (t) - E[x(t) x T (t)] H T (t) = E[(x(t) - x(t)) x T (t)] H T (t) = {E[x(t) x T (t)] + E[x(t) x T (t)]} H T (t) 59 The condition that x is the best estimate from among all linear func- T tionals of z(t) leads to the result that E[x(t) x (t)] = 0. Hence P(t) H T (t) = E[x(t) x T (t)] H T (t) = K(t, t) R v (t) (5.12) From the stochastic integral, r x(t) = J K(t, r) dz(r) r H T R _1 dz(t) + (A(t) K(t, r) - K(t, t) H(t) K(t, r)) dz(r) dt o : -(\ Mi ) ,I:..U) • i 9 K( ^' r) dz(r) dt 3t = P dx(t) = A(t) x(t) dt + P H T R 1 (dz(t) - H(t) x(t) dt) (5.13) Since z(t) and presumably dz(t) are known, this is a stochastic differ- ential equation for x(t). Note that dz(t) - H(t) x(t) dt = dz(t) - H(t) x(t) dt + H(t) x(t) dt = dv(t) + H(t) x(t) dt From this expression and (5.13), we get the following stochastic differential equation for x 60 dx = dx(t) - dx(t) T -1 =Axdt+dw-Axdt-P H R (dv + H(t) x) = A x dt + dw - P H T R dv - P H T R H x dt = [A - P H T R 1 H] x dt + dw - P H T R 1 dv with 5(0) = x(O). By the methods developed in the previous section for stochastic differential equations, P(t) = E[x(t) x T (t)] 1 J $(t, a) = $(t, 0) r $ (t, 0) + $(t, a) [Q(a) + P(o) H T (a) R 1 (a) H(a) P T ] $ T (t, a) da (5.14) where $(t, t) is the transport matrix associated with the linear differential equation 4*- = (A - P H T R _1 H) y (5.15) dt v and where from. (4.11) Q dT = E[dw(x) dw T (x)] (5.16) 4^- = (A - P H T R X H) P + P(A - P H T R _1 H) T dt v v T -1 T + Q(t) + P H R H P 61 dP dt = A P + P A T - T -1 P H R H P - Q (5.17) p(0) = r (5.18) This set of equations finishes the solution of the filter problem. The optimal filter is a feedback system which is described by the stochastic differential equation (5.13). It is obtained by taking the measurements z(t), forming the error signal z(t) - H(t) x(t) , and feed- T -1 ing the error forward with a gain P(t) H (t) R (t) . P(t), the error variance, is obtained as a solution to the nonlinear Riccati-type equation (5.17), H(t) is a known transformation matrix, and R is v the variance of the Wiener process dv. A block diagram of the filter is shown in Figure 8. The variables appearing in this diagram are vectors, and the boxes represent matrices operating on vectors. The double lines which indicate direction of signal flow serve as a reminder that multiple signals rather than a single one are being directed. Figure 8 — Optimal Filter 62 REFERENCES 1. Astrom, K. J., "Introduction to Stochastic Control Theory," Academic Press, Inc., New York (1970). 2. Kalman, R. E. and R. S. Bucy, "New Results in Linear Filtering and Prediction Theory," Journal of Basic Engineering Series D, American Society of Mechanical Engineers, Vol. 83, pp. 95-108 (1961). 3. Bucy, R. S. and P. D. Joseph, "Filtering for Stochastic Processes with Applications to Guidance," Interscience Publishers, Inc., New York (1968). 4. Bellman, R. E. and S. E. Dreyfus, "Applied Dynamic Programming, Princeton University Press, Princeton, N.J. (1962). 5. Coddington, E. A. and N. Levinson, "Theory of Ordinary Differ- ential Equations," McGraw-Hill Book Company, Inc., New York (1955). 6. Bailey, P. B., et al. , "Nonlinear Two Point Boundary Value Problems," Academic Press, Inc., New York (1968). 7. Dreyfus, S. E., "Dynamic Programming and the Calculus of Variation," Academic Press, Inc., New York (1965). 8. Larson, R. E. , "State Increment Dynamic Programming," American Elsevier Publishing Company, Inc., New York (1968). 9. Doob, J. L. , "Stochastic Processes," Wiley, Inc., New York (1963). 10. Gikhman, I. I. and A. V. Skorokhod, "Introduction to the Theory of Random Processes," W. B. Saunders Company, Philadelphia (1969). 63 Copies 1 US Army Waterways Experiment Station Res Center Lib 1 CHONR R. D. Cooper, Code 438 1 ONR BOSTON 1 ONR CHICAGO 1 ONR PASADENA 1 USNA P. Van Mater 1 NAVPGSCOL NROTC & NAVADMINU 1 NAVWARCOL 1 NRL/Lib 4 NAVSEA 2 SEA 09G32 1 SEA 0331G 1 SEA 035B 1 NAVFACENGCOM 1 NAVOCEANO/Lib 1 NADC 1 NELC 1 NWC 1 NUC, San Diego 1 NUC, Pasadena 1 NCEL/Code L31 1 NSWC, White Oak/Lib 1 NSWC, Dahlgren/Lib INITIAL DISTRIBUTION Copies 1 NUSC NPT 1 NUSC NLONLAB 1 NAVSHIPYD BREM/Lib 1 NAVSHIPYD CHASN/Lib 1 NAVSHIPYD MARE/Lib 1 NAVSHIPYD NORVA/Lib 1 NAVSHIPYD PEARL/Lib 1 NAVSHIPYD PHILA/Lib 1 NAVSHIPYD PTSMH/Lib 10 NAVSEC 1 SEC 6034B 2 SEC 6110 1 SEC 6114D 1 SEC 6114H 1 SEC 6120 1 SEC 6136 1 SEC 6140B 1 SEC 6144G 1 SEC 6660.03/D.L. Blount 1 AFFDL/FDDS/J. Olsen 2 AFFDL/FYS 1 Dale Cooley 1 S.J. Pollock 12 DDC 2 COGARD 1 COM (E) , STA 5-2 1 Diy of Merchant Marine Safety 1 LC/SCI & TECH DIV 1 MARAD/Adv Ship Prog Off 1 MMA/Tech Lib 1 NASA AMES RES CEN R.T. Medan, MS 221-2 65 Copies Copies 3 NASA LANGLEY RES CEN 1 J.E. Lamar, Ms 404A 1 Brooks 1 E.C. Yates, Jr., Ms 340 1 NASA Sci & Tech Info Facility 1 NSF/Eng Div 1 Univ of Bridgeport Prof E. Uram Mech Eng Dept 4 Univ of California, Berkeley College of Eng, NA Dept 1 Lib 1 J.R. Paulling 1 J.V. Wehausen 1 H.A. Schade 3 Calif Inst of Tech 1 A.J. Acosta 1 T.Y. Wu 1 M.S. Plesset 1 Colorado State Univ M. Albertson Dept of Civ Eng 1 Univ of Connecticut V. Scottron Hyd Res Lab 1 Cornell Univ/W.R. Sears Grad School of Aero Eng 1 Florida Atlantic Univ Ocean Eng Lib 1 Harvard Univ/Dept of Math G. Birkhoff 1 Univ of Hawaii Dr. Bretschneider 3 State Univ of Iowa Iowa Inst of Hyd Res 1 L. Landweber 1 J. Kennedy 1 Hunter Rouse 1 Kansas State Univ Engineering Exp Station D.A. Nesmith 1 Lehigh Univ/Fritz Lab Lib 1 Long Island Univ Grad Dept of Marine Sci David Price 1 Delaware Univ/Math Dept 4 Univ of Maryland 1 Eng Lib 1 P.F. Cunniff 1 C.L. Sayre 1 F. Buckley 6 Mass Inst of Technol Dept of Ocean Eng 1 P. Mandel 1 J.R. Kerwin 1 N. Neumann 1 P. Leehey 1 M. Abkowitz Hydro Lab 1 A.T. Ippen 3 Univ of Mich/Dept NAME 1 T.F. Ogilvie 1 H. Benford 1 R.B. Couch 5 Univ of Minn/ St Anthony Falls 1 C.S. Song 1 J.M. Killen 1 F. Schiebe 1 J.M. Wetzel Univ of Illinois College of Eng J.M. Robertson Theoretical & Applied Mech New York Univ 1 W.J. Pierson, Jr. Courant Inst of Math Sci 1 A.S. Peters 1 J.J. Stoker 66 Copies 1 Univ of Notre Dame A.F. Strandhagen 1 Penn State Univ Ordnance Res Lab 1 St. John's Univ/Math Dept Jerome Lurye 3 Southwest Res Inst 1 H.N. Abramson 1 G.E. Tansleben, Jr. 1 Applied Mech Review 3 Stanford Univ/Dept of Civ Eng 1 R.L. Street 1 B. Perry Dept of Aero and Astro 1 H. Ashley 1 Stanford Res Inst/Lib 3 Stevens Inst of Tech Davidson Lab 1 J. P. Breslin 1 S. Tsakonas 1 Lib 1 Utah State Univ/Col of Eng Roland W. Jeppson 2 Univ of Virginia/Aero Eng Dept 1 J.K. Haviland 1 Young Yoo 2 Webb Institute 1 E.V. Lewis 1 L.W. Ward 1 Worcester Poly Inst/Alden Res Lab 1 Woods Hole, Ocean Eng Dept 1 SNAME 1 Aerojet-General W.C. Beckwith 1 Bethlehem Steel Sparrows A.D. Haff, Tech Mgr Copies 1 Bolt, Beranek & Newman, MA 11 Boeing Company/Aerospace Group 1 R.R. Barbar 1 H. French 1 R. Hatte 1 R. Hubard 1 F.B. Watson 1 W.S. Rowe 1 T.G.B. Marvin 1 C.T. Ray Commercial Airplane Group 1 Paul E. Rubber t 1 Gary R. Saaris 1 Cornell Aero Lab Applied Mech Dept 1 Flow Research, Inc Frank Dvorak 1 Eastern Res Group 2 General Dynamics Corp 1 Convair Aerospace Div A.M. Cunningham, Jr. MS 2851 1 Electric Boat Div V.T. Boatwright, Jr. 1 Gibbs & Cox, Inc. Tech Info Control Section 1 Grumman Aircraft Eng Corp W.P. Carl, Mgr, Grumman Marine 1 S.F. Hoerner 2 Hydronautics, Inc. 1 P. Eisenberg 1 M.P. Tulin 4 Lockheed Aircraft Corp Lockheed Missiles & Space 1 R.L. Waid 1 R. Lacy 1 Robert Perkins 1 Ray Kramer 1 Marquardt Corp/F. Lane General Applied Sci Labs 67 Copies 1 Copies Code Martin Marietta Corp/Rias Peter F. Jordan McDonnell-Douglas Corp Douglas Aircraft Company 1 A.M.O. Smith 1 Joseph P. Giesing 1 Newport News Shipbuilding/Lib 1 Nielsen, NA Rockwell 1 North American Rockwell Los Angeles Div Jan R. Tulinius Dept 056-015 2 Northrop Corp Aircraft Div 1 J.T. Gallagher 1 J.R. Stevens 1 Oceanics, Inc. Paul Kaplan 1 Sperry Sys Mgmt 1 Robert Taggart, Inc. 1 Tracor CENTER DISTRIBUTION Copies Code 1 11 1 115 1 1151 1 1152 1 1154 1 15 1 1502 1 1504 1 1505 1 1506 1 1507 1 152 1 1521 152 1528 1532 Wen Lin 1 154 1 1541 1 1542 1 1544 1 1548 3 1552 1 J. McCarthy 1 K.P. Kerney 50 T. J. Langan 1 H.T. Wang 3 1556 1 P.K. Besch 1 E.P. Rood 1 D. Coder 1 156 1 1568 2 1572 1 M. Ochi 1 ( :. Lee 1 1576 1 16 1 167 1 169 1 I. J. Engler 17 1 18 2 1843 1 19 1 1966 1 273 1 2732 30 5211 1 5221 1 5222 Y. Liu Report Distribution Library (C) Library (A) 68